Abstract: Abstract：In view of the lack of remote sensing scene data, the obvious size change of surface objects captured by aerial photography, including a large number of objects of multiple categories and complex background, resulting in low detection accuracy and inaccurate classification, a small sample remote sensing target detection network based on the two-stage detection model (Faster RCNN) is proposed. New involution convolution operators are added to build detector backbone to improve feature extraction capability; Integrate multi-scale object-level positive sample features to enhance the original features, suppress the adverse effects of negative samples, fully mine the feature information of each target scale, and help the semantic information to locate; The idea of comparative supervision is adopted to improve the loss function, refine the target classification and reduce the false detection rate. The experimental results on public remote sensing data sets show that the network can adapt to the multi-scale characteristics of remote sensing images and effectively alleviate the over-fitting phenomenon caused by data scarcity under the condition of only a small number of remote sensing labeled samples. Compared with the previous Meta RCNN and FsDet networks, the average accuracy has been further improved by 3.8 percentage points and 2.5 percentage points, providing a meaningful reference for image target detection in the remote sensing field.

Key words: Key words: few shot, object detection, feature enhancement, fine tuning, remote sensing images, contrastive loss